DexVLA
VLA with a 1B-parameter plug-in diffusion expert and a three-stage curriculum, built for dexterous multi-embodiment control.
About
DexVLA (February 2025, arXiv 2502.05855) argues the bottleneck in VLAs is not the vision-language backbone but the action head. Its answer is a 1B-parameter diffusion-based action expert trained separately on cross-embodiment data, then plugged into a Qwen2-VL (~2B) backbone. Training runs in three stages: pre-train the diffusion expert, align the VLM to specific embodiments, then post-train for fast adaptation. It was tested on single-arm, bimanual, and dexterous-hand setups, and the authors report it beating Octo, OpenVLA, and Diffusion Policy across the board — including a 0.92 score on shirt folding with no task-specific adaptation, a task the baselines could not complete. The code is public at github.com/juruobenruo/DexVLA and there are community adaptations (for example a UR10 port), which is decent evidence the release is real and usable.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- Paper (arXiv 2502.05855): beats Octo, OpenVLA, Diffusion Policy on single-arm, bimanual, and dexterous tasks; 0.92 on shirt folding zero-shot (author-reported)
- No independent benchmark replication found
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 8/10
Scored 8/10 (TIER verified, metrics-v1). Evidence: Midea/ECNU paper (arXiv 2502.05855) with evaluation across single-arm, bimanual, and dexterous-hand embodiments; public code repo with community adaptations (e.g. UR10 port). Strengths: novel separable diffusion-expert design, strong reported results on hard long-horizon tasks, usable release. Docked for: (1) benchmarks are author-reported; (2) license not confirmed from repo metadata at research time. License: unverified — check repo before commercial use.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 7.0/10
- Maintenance · recency, releases, issue load
- 6.0/10
- Openness · license + docs
- 9.0/10
- Manual review · sanity read + red-flag check
- 8.0/10
- · Midea Group / ECNU paper arXiv 2502.05855 (Feb 2025); 1B diffusion action expert + Qwen2-VL ~2B backbone
- · Three-stage curriculum: pretrain expert, align VLM, post-train for adaptation
- · Community forks exist (e.g. UR10 adaptation) — usable public release